"PyTorch" matching MCP servers:

A full-featured MCP server for PyTorch documentation workflows, providing tools for search, symbol lookup, code examples, troubleshooting, and question-answering using local docs.

Gives Claude Code, Cursor and other agents structural awareness of a PyTorch model: layers, params, FLOPs, blast radius, the design linter, a full readiness/cost/deployment verdict, and a ranker for which of k candidate designs to train. Reads a .py, a .neurarch.json, 81 bundled reference architectures, or a Hugging Face repo. Offline, no API key.